SLB API Console

Graph Extraction Pipeline — Quick Map

Follow these steps in order to go from raw report → ontology → graph → Q&A.

Upload & Extract

Convert the report into clean text + metadata.

→

Ontology Extraction: adversarial step

Generate initial classes + relationships from the text.

→

Ontology Lookup

Match extracted terms to the reference TTL in Neo4j.

→

Candidate Creation

Propose missing categories/relations for review.

→

Candidate Decision

Approve or reject the candidate changes.

→

Graph Extraction

Build validated triples (optionally push to Neo4j).

→

Q&A Chat

Ask questions and explore the report insights.

1Content extraction

/v1/extract

Step 1 — Upload a document to extract clean text + metadata.

What you get

Use this first to ingest a report before running ontology or graph steps.

Accepted files: PDF, DOCX, PPTX, XLSX, TXT.
Output: extracted text, detected language, pages/tables info.

23Ontology tools

/v1/fa/extract-ontology · /v1/fa/lookup-ontology

Step 2 — Convert text into an ontology, then match it to the reference TTL.

Why this matters

This block answers: “What entities + relationships exist in this report?”

Extraction = propose classes/relations from text.
Lookup = match those classes to the reference TTL in Neo4j.

2Ontology extraction: adversarial step

/v1/fa/extract-ontology

Build a compact list of classes + relationships from the report text. Create a critical version of it from adversarial LLM and refine it until completion.

Document type (controls expectations) Used to decide which expected categories are relevant. Expected categories (comma-separated) Comma-separated list of categories you want to see if possible. Max categories (cap) Upper limit to keep the ontology compact. Max relationships (cap) Upper limit for relationships (prevents over-generation). Text

3Ontology lookup

/v1/fa/lookup-ontology

Compare the extracted ontology to the reference TTL (Neo4j).

Ontology JSON (from extraction)

{
  "categories": [
    "FA Report",
    "Part",
    "Test",
    "Observation",
    "Failure Mechanism",
    "Root Cause",
    "Recommendation",
    "Material",
    "Test Method"
  ],
  "relationshipTypes": [
    {
      "name": "documents",
      "fromCat": "FA Report",
      "to": "Part"
    },
    {
      "name": "includes_test",
      "fromCat": "FA Report",
      "to": "Test"
    },
    {
      "name": "records_observation",
      "fromCat": "FA Report",
      "to": "Observation"
    },
    {
      "name": "identifies_mechanism",
      "fromCat": "FA Report",
      "to": "Failure Mechanism"
    },
    {
      "name": "identifies_root_cause",
      "fromCat": "FA Report",
      "to": "Root Cause"
    },
    {
      "name": "provides_recommendation",
      "fromCat": "FA Report",
      "to": "Recommendation"
    },
    {
      "name": "uses_material",
      "fromCat": "Part",
      "to": "Material"
    },
    {
      "name": "applies_test_method",
      "fromCat": "Test",
      "to": "Test Method"
    }
  ],
  "properties": [
    {
      "name": "description",
      "domain": "Observation"
    },
    {
      "name": "description",
      "domain": "Root Cause"
    },
    {
      "name": "action",
      "domain": "Recommendation"
    },
    {
      "name": "material_type",
      "domain": "Material"
    },
    {
      "name": "method",
      "domain": "Test Method"
    }
  ]
}

Paste the JSON output from extraction (or build your own).

45Ontology candidates

/v1/fa/candidate-ontology

Step 3 — Propose missing categories if coverage is too low.

When to use

If your extracted ontology misses key categories, you can propose a candidate.

Use this when coverage is below your threshold.
Approving a candidate can update the reference TTL.

4Create a candidate

/v1/fa/candidate-ontology

Create a candidate for missing classes/relations.

Text coverage (0–1) How much of the ontology was matched (0–1). Missing categories (comma-separated) Comma-separated labels that were not matched. Missing relationships (JSON) Only include relationships justified by the text.

5Approve a candidate

/v1/fa/candidate-ontology/<id>/(approve|reject)

Approve or reject a candidate and merge into the TTL.

1-6Graph Extraction (full pipeline)

/v1/fa/run-workflow

End-to-end graph extraction pipeline with SME validation.

End-to-end tutorial

Use this when you want the full graph extraction pipeline end-to-end.

Phase 1 runs extract → ontology → lookup → graph automatically when coverage is sufficient.
If it returns pending_sme, a candidate needs validation.
Phase 2 approves/rejects the candidate, then resumes and extracts the graph.

Phase 1 — Extraction & validation

upload → (done | pending_sme)

Runs the pipeline end-to-end (extract → ontology → lookup → graph) unless an ontology candidate must be validated. If it returns pending_sme, run phase 2 to approve/reject the candidate and resume graph extraction.

File (FA report) Upload the FA report (PDF/DOCX/etc.). UUID is assigned automatically. VLM description for tables VLM description for images Document type Controls expected categories and behavior. Min coverage threshold If coverage falls below this, a candidate is created. Max chunks (context window) Higher values improve context, but cost more.

Phase 2 — Decision & graph

approve → extract graph

Validates the candidate and extracts the final graph.

Candidate ID (from phase 1) Candidate created during phase 1. Decision Approve to merge, reject to stop. Decided by (SME) SME username or role. Comment (optional) Optional reason for traceability. Max chunks (context window) Text chunks used for graph extraction.

7 Q&A Chat

Sign in